Workflow Modes

PAW supports three workflow modes to match your task scope and development style. Each mode determines which stages are included and how work is reviewed.

Quick Overview

Mode	Stages	Review Strategy	Best For
Full	All stages (Spec → Docs)	PRs or Local	Large features, complex changes
Minimal	Core stages (Code Research → Final PR)	Local only	Bug fixes, small enhancements
Custom	User-defined	PRs or Local	Unique workflows

Full Mode

Stages Included: Spec → Spec Research → Code Research → Implementation Plan → Implementation (including Documentation phase) → Final Review (if enabled) → Final PR → Status

Full mode is the complete PAW workflow with all stages from initial specification through comprehensive documentation. When Final Agent Review is enabled (default), an automated review step runs after implementation phases to catch issues before the Final PR.

When to Use Full Mode

• Large features requiring comprehensive documentation
• Complex system changes needing formal specifications
• New features where requirements need refinement through the spec process
• Work that benefits from detailed technical documentation for future maintainers
• Projects where complete traceability and documentation are organizational requirements

Review Strategies for Full Mode

Full mode supports both review strategies:

• PRs strategy: Creates intermediate PRs (planning, phase, docs branches) for review at multiple stages. Best for complex work requiring review checkpoints throughout development.

• Local strategy: Single branch workflow with all work on the target branch, only creating the final PR. Best when you prefer consolidated review but still want all stages executed.

Minimal Mode

Stages Included: Code Research → Implementation Plan → Implementation (including Documentation phase) → Final Review (if enabled) → Final PR → Status

Stages Skipped: Specification stage

Minimal mode is a streamlined workflow focusing on core implementation activities. It assumes requirements are already clear from the issue description or enhancement request.

When to Use Minimal Mode

• Bug fixes with clear requirements from issue description
• Small features or enhancements where the goal is well-understood
• Refactoring work where objectives and approach are already clear
• Quick iterations when formal specs aren't needed
• Work where existing project documentation is sufficient

Review Strategy for Minimal Mode

Minimal mode enforces local strategy (single branch workflow) to simplify the process. No intermediate planning, phase, or docs branches or PRs—all work happens on the target branch with only a final PR created.

Quality Assurance

Even though specification and documentation stages are skipped, all quality gates (tests, linting, type checking, build verification) remain mandatory. The implementation plan still includes detailed success criteria and phase breakdown for reviewable work.

Custom Mode

Stages Included: User-defined based on custom instructions provided during initialization

Custom mode offers flexibility to define which stages to include and how work should be reviewed. Agents interpret your custom instructions to determine the appropriate workflow structure.

When to Use Custom Mode

• Unique workflows that don't fit full or minimal patterns
• Experimenting with different workflow configurations
• Project-specific requirements that need non-standard stage combinations
• Workflows where you want some but not all optional stages

Review Strategies for Custom Mode

Custom mode supports both prs and local strategies. You specify the review approach as part of your custom instructions.

Example Custom Instructions

Skip specification stage, include documentation, use prs review strategy

Include code research and implementation only, single branch workflow

Full stages but combine all implementation phases into one, use local strategy

Review Strategies Explained

PRs Strategy (Intermediate Pull Requests)

Branch Structure:

• Planning branch: <target>_plan → Planning PR to target branch
• Phase branches: <target>_phase[N] → Phase PRs to target branch
• Docs branch: <target>_docs → Docs PR to target branch
• Final PR: target branch → base branch (usually main)

Best for:

• Complex work requiring review at multiple stages
• Large features where incremental review reduces cognitive load
• Teams with multiple reviewers who can review different phases
• Work where early feedback on planning or individual phases is valuable
• Projects where intermediate review checkpoints improve quality

Trade-offs:

• More PR overhead (planning, phases, docs, final)
• Longer time from start to merge (multiple review cycles)
• Better quality assurance through staged review
• Easier to rewind and fix issues at specific stages

Local Strategy (Single Branch)

Branch Structure:

• All work committed directly to target branch
• Final PR: target branch → base branch (usually main)

Best for:

• Simpler work where consolidated review is preferred
• Solo developers or small teams with streamlined review processes
• Work where you want to complete all stages but prefer single review point
• Bug fixes or small enhancements (especially with minimal mode)
• Projects where PR overhead should be minimized

Trade-offs:

• Less PR overhead (only final PR)
• Faster path to merge (single review cycle)
• Larger final PR to review (all stages combined)
• Harder to rewind to specific stages (requires git operations on target branch)

Quality Gates

Important

Quality gates remain mandatory regardless of workflow mode or review strategy. Skipping stages or choosing local strategy streamlines the process but never compromises code quality.

All automated verification criteria in implementation plans must pass before work can proceed, regardless of which workflow mode is selected:

• Tests must pass
• Linting must pass
• Type checking must pass
• Build verification must pass

Final Agent Review Configuration

Final Agent Review runs after all implementation phases complete, before the Final PR is created. It catches issues before external review. Configuration is stored in WorkflowContext.md.

Field	Default	Options	Description
Final Agent Review	enabled	enabled, disabled	Whether to run the review step
Final Review Mode	multi-model	single-model, multi-model, society-of-thought	Review mode (see Society-of-Thought Review)
Final Review Interactive	smart	true, false, smart	true: confirm each finding; false: auto-apply; smart: auto-apply consensus, interactive for ambiguous
Final Review Models	latest GPT, latest Gemini, latest Claude Opus	comma-separated	Models for multi-model review (CLI only)
Final Review Specialists	all	all, comma-separated names, adaptive:<N>	Which specialists participate (society-of-thought only)
Final Review Interaction Mode	parallel	parallel, debate	How specialists interact (society-of-thought only)

Notes:

• VS Code only supports single-model mode (multi-model and society-of-thought fall back to single-model)
• CLI supports all three modes; multi-model spawns parallel reviews and synthesizes findings
• In smart mode (default), consensus fixes are auto-applied and ambiguous findings prompt for your decision
• In interactive mode, you confirm each finding before changes are made
• In auto-apply mode, must-fix and should-fix recommendations are applied automatically

Multi-Model Planning Configuration

Planning can use multiple AI models to independently create competing implementation plans, then synthesize the best elements. This brings the multi-model pattern to where it has the highest leverage — design decisions.

Field	Default	Options	Description
Plan Generation Mode	single-model	single-model, multi-model	Plan generation execution mode
Plan Generation Models	latest GPT, latest Gemini, latest Claude Opus	comma-separated	Models for multi-model plan generation (CLI only)

Modes:

• single-model — Current behavior. One model creates the plan. (Default)
• multi-model — Three models independently create plans in parallel, then a synthesis step merges them. (N+1 LLM calls)

Notes:

• Multi-model planning is CLI only — VS Code falls back to single-model
• Per-model artifacts are saved in a gitignored planning/ subfolder (distinct from reviews/ which holds final review artifacts)
• With all multi-model options enabled (plan generation + final review), expect 8 LLM calls before any code is written. This is the max-quality configuration, not the default.

Planning Docs Review Configuration

Planning Docs Review runs during planning (Stage 02) to validate cross-artifact consistency before implementation begins. Configuration is stored in WorkflowContext.md.

Field	Default	Options	Description
Planning Docs Review	enabled	enabled, disabled	Whether to run the review step
Planning Review Mode	multi-model	single-model, multi-model	Single model or parallel multi-model review
Planning Review Interactive	smart	true, false, smart	true: confirm each finding; false: auto-apply; smart: auto-apply consensus, interactive for ambiguous
Planning Review Models	latest GPT, latest Gemini, latest Claude Opus	comma-separated	Models for multi-model review (CLI only)

Selecting Your Mode

When using the VS Code extension's PAW: New PAW Workflow command, you'll be prompted to:

1. Select workflow mode (Full, Minimal, or Custom)
2. Select review strategy (PRs or Local)—automatically set to Local for Minimal mode
3. Provide custom instructions (if Custom mode selected)

Your selections are stored in WorkflowContext.md and guide all agents throughout the workflow.